On scales and dynamics in observing the environment

Abstract

Natural and anthropogenic processes at the Earth's surface operate at a range of spatial and temporal scales. Different scales of observation are required to match the spatial scales of the processes under observation. At the same time, the temporal sampling rate of the observing systems must be reconciled with the dynamics of the processes observed. Bringing together these issues requires insight, innovation and, inevitably, compromise. This paper reviews spatial and temporal considerations in remote sensing and introduces the papers in this Special Issue on ‘Scales and Dynamics in Observing the Environment’. The review comprises three main sections. The first section focuses on spatial variability in remote sensing, while the second section focuses on temporal variability in remote sensing. The third section links these two issues, focusing on the interplay of space and time in remote sensing. The review is primarily theoretical, explaining spatial and temporal properties of remote sensing and remotely sensed phenomena. Where appropriate, however, practical examples are included to demonstrate how remote sensing is used in environmental applications. Following the review, the papers included in the Special Issue are introduced, outlining their significance in the context of ‘Scales and Dynamics in Observing the Environment’.

1. Introduction

Natural and anthropogenic processes at the Earth's surface operate at a range of spatial and temporal scales. Different scales of observation are required to match the spatial scales of the processes under observation. At the same time, the temporal sampling rate of the observing systems must be reconciled with the dynamics of the processes observed. Bringing together these issues requires insight, innovation and, inevitably, compromise.

This opening paper serves as both a review of spatial and temporal considerations in remote sensing and an introduction to the papers in this Special Issue on Scales and Dynamics in Observing the Environment. The review comprises three main sections. The initial section focuses on spatial variability in remote sensing, followed by a section on temporal variability in remote sensing. The next section links these two issues, focusing on the interplay of space and time in remote sensing. Following the review, the papers included in the Special Issue are introduced, outlining their significance in the context of ‘Scales and Dynamics in Observing the Environment’. Clearly, this overall theme is relatively broad. As such, it should be noted that the intention of this paper, and more generally the Special Issue, is to provide a summary of key issues and significant developments, rather than a comprehensive treatment of the subject.

2. Spatial variability

The spatial variability of the Earth's surface is represented in remote sensing according to certain spatial properties of the imagery, in particular the size and nature of, and relationships between, pixels. The most significant spatial element of remotely sensed imagery is the pixel size, which is related closely, although not synonymously, with the more general term of spatial resolution. In fact, no single quantitative definition of spatial resolution exists. Several definitions have been proposed, based on, for example, the geometrical properties of the sensor (such as instantaneous field of view), the ability to separate point targets (point spread function), the ability to measure periodicity of repetitive targets (modulation transfer function) and the ability to measure spectral properties of small objects (effective resolution element) (Townshend 1981, Forshaw et al. 1983, Cracknell 1998, Fisher 1998). In general terms, spatial resolution refers to the spatial detail of remotely sensed imagery. Imagery at finer spatial resolution provides greater spatial detail and, therefore, enables the delineation of smaller features than coarser spatial resolution imagery.

Strahler et al. (1986) introduced the concept of the L‐ and H‐models of spatial resolution, a concept that influenced much subsequent work (e.g. Jupp et al. 1988, Raffy 1992), and continues to do so (e.g. Wulder et al. 2004, Prenzel and Treitz 2005). In the L‐resolution case, the variability of the Earth's surface is at a higher spatial frequency than image sampling (surface features of interest are smaller than the spatial resolution of the imagery), such that surface features cannot be resolved spatially. In the H‐resolution case, the Earth's surface varies at a lower spatial frequency than image sampling and features can be resolved. L‐resolution is the cause of a common problem in identifying features from remotely sensed imagery, that of mixed pixels, so called because they represent areas comprising a mixture of two or more features, such as land‐cover types (Chhikara 1984, Cracknell 1998, Naceur et al. 2004). This can result in error where pure land‐cover types are the focus of analysis, as is usually the case. A common cause of mixed pixels is the presence of land‐cover boundaries, where pixels represent a weighted average of the spectral response of the land‐cover types on each side of the boundary. In general, for any specific spatial frequency of land cover, at a fixed level of classification generalization, the proportion of mixed pixels is likely to increase as spatial resolution becomes coarser.

Since relatively fine spatial resolution imagery tends to provide more spatial detail and, proportionally, fewer mixed pixels, than coarser spatial resolution imagery, it would seem beneficial to use the former where possible. However, this is not always the case. Where H‐resolution exists, increasing the spatial resolution may lead to oversampling, resulting in variation within features. Such variation can lead to error in feature identification (Cushnie 1987, Hsieh et al. 2001, Aplin and Atkinson 2004). For example, where the feature of interest is a forest stand but the spatial resolution of the imagery is sufficiently fine to distinguish between individual trees and patches of grass separating trees, certain pixels may be misidentified as grass. Under certain circumstances, therefore, it may be beneficial to identify features using coarser, rather than finer, spatial resolution imagery (Townshend and Justice 1988, Pax‐Lenney and Woodcock 1997).

The spatial resolution of remotely sensed imagery is only a limiting factor where it is too coarse to enable the delineation of required features. Imagery with a spatial resolution that is too fine for the required purpose can be degraded through image resampling or averaging to any chosen (coarser) spatial resolution (Justice et al. 1989, Kumar et al. 1996, Ju et al. 2005). In contrast, there is no generally applicable procedure to increase (i.e. make finer) the spatial resolution of imagery that is too coarse to meet requirements, although several methods have been developed that approach this problem. Initially, much work was focused on characterizing the relative constituents contributing to mixed pixels in a proportional sense. That is, where mixed pixels comprise various different features or types of features (e.g. land‐cover classes), techniques such as mixture modelling (Ashton and Schaum 1998, Liu and Wu 2005), fuzzy classification (Wang 1990, Shalan et al. 2003) or linear regression (Oleson et al. 1995) were developed to identify how much of each feature was present. For example, a pixel positioned on the boundary between woodland and grassland may comprise 60% woodland and 40% grassland. However, although able to identify the ‘amount’ of each feature within each pixel, these proportional techniques are unable to indicate the ‘location’ of these features within the pixel.

Recently, techniques have been developed that characterize the relative constituents contributing to mixed pixels in a spatial sense. For example, Steinwendner et al. (1998) employed a simple moving window approach to locate linear features on a sub‐pixel basis, Aplin and Atkinson (2001) used vector data to guide the allocation of sub‐pixel land‐cover proportions from a fuzzy classified image, Verhoeye and De Wulf (2002) exploited the spatial relationships between neighbouring pixels to map sub‐pixel land cover, while Mertens et al. (2004) used artificial neural networks and wavelets for sub‐pixel mapping and sub‐pixel sharpening. Perhaps the most promising such approach, however, is ‘super‐resolution’ mapping, that uses an iterative pixel swapping algorithm to exploit the spatial dependencies between sub‐pixels, yielding highly accurate sub‐pixel land‐cover classifications (Tatem et al. 2002, Atkinson 2005, Kasetkasem et al. 2005).

When considering spatial resolution, it is important to realize that features on the Earth's surface, as identified by imagery with a specific spatial resolution, may be scale‐dependent (Turner et al. 1989, Walsh et al., 1999, Asner et al. 2003). That is, features may be identified relatively accurately at one spatial resolution, but largely misidentified at another. Much research has been conducted on identifying the ‘optimal’ spatial resolution for a particular purpose or investigation, where there is a prior assumption that the features under investigation are scale‐dependent. For example, Welch (1982) simply used the minimum mappable unit of urban land cover to estimate the most appropriate spatial resolution. Alternatively, images produced by sensors with different spatial resolutions have been compared to establish which is most suitable for a particular purpose (e.g. Nelson and Holben 1986). Developing a more complex approach, Woodcock and Strahler (1987) degraded a relatively fine spatial resolution image to successively coarser spatial resolutions to identify the most appropriate scale of observation. Many researchers have since used image degradation as a means of investigating scale and establishing a suitable spatial resolution (Justice et al. 1989, Kumar et al. 1996, Chen et al. 2004). However, it is perhaps the field of geostatistics that holds the greatest potential for characterizing optimal scales in remote sensing (Curran and Atkinson 1998). In addition to techniques such as correlograms and clustering (Legendre and Fortin 1989, Cullinan and Thomas 1992), there has been particular interest in using variograms (Atkinson and Curran 1997, Treitz 2001) and local variance (Woodcock and Strahler 1987, Rahman et al. 2003, Emerson et al. 2005) for this purpose.

In contrast to work on identifying the optimal spatial resolution, researchers have recently questioned whether it is, in fact, appropriate to select a single spatial resolution for a remotely sensed image (Atkinson and Aplin 2004). For example, Ju et al. (2005), referring to the example of land‐cover classification, contend that in a complex image, a single scale may not adequately represent all the classes of interest. An appropriate alternative to single‐scale observations may be to use multi‐scale analysis (Coops and Waring 2001, Andrefouet et al. 2003, Li et al. 2003, Katartzis et al. 2005). In simple terms, multi‐scale remote sensing analysis involves the observation of the Earth's surface at more than one spatial scale. Where features are scale‐dependent, they can be represented differently at different spatial scales or resolutions. Therefore, observation at multiple spatial resolutions can increase the volume of information available to characterize and distinguish features.

One relatively common form of multi‐scale analysis is upscaling, whereby detailed local measurements are extrapolated over wider areas (Hay et al. 1997, Anderson et al. 2004, Wang et al. 2004). Conversely, as da Silva et al. (2005) demonstrate, it is also possible to ‘downscale’ and infer local detail from wider‐scale information. Where distributions of features are examined over a range of scales, careful consideration should be given to how feature types will be affected by varying spatial resolution (Moody and Woodcock 1995, Raffy and Blamont 2003, Wang et al. 2004). For example, using the example of land‐cover classification, it may not be appropriate to maintain identical feature classes at different spatial resolutions. Instead, it may be more appropriate to redefine classes. At a spatial resolution of 1 km, urban areas may be distinguished accurately as a single ‘urban’ class since the pixels are likely to represent an average of all the features within the urban area. That is, urban areas at this scale would probably have a relatively constant spectral response. However, the same class may be less accurately classified using 10 m spatial resolution imagery since the pixels are likely to represent individual features such as buildings, lawns and roads. Individual pixels within urban areas, therefore, may be misclassified as other classes. To classify a single ‘urban’ class accurately using 10 m spatial resolution imagery, it may be necessary to classify individual sub‐classes, such as roof tiles, grassland and asphalt, and subsequently merge these sub‐classes to form a general class.

Spatial scale and resolution have received a great deal of attention from the remote sensing community and a considerable body of work has been devoted to investigating how the scale of observation affects the representation of features on the Earth's surface. As noted above, in essence, spatial resolution is a fairly simple concept (finer spatial resolution imagery enables the delineation of smaller features than coarser spatial resolution imagery), yet many uncertainties and difficulties remain regarding the issue of scale in remote sensing. For example, Strahler et al. (1986) introduced the L‐ and H‐models of spatial resolution, characterizing the problems of mixed pixels and within‐feature variability, respectively. As such, the choice of spatial resolution would seem key to success in remote sensing studies. This may be true under certain circumstances, but Atkinson and Aplin (2004) and Ju et al. (2005) question whether selecting a single spatial resolution is appropriate at all, and suggest that multi‐scale investigation may be more appropriate. Overall, therefore, it may be inappropriate to make overarching recommendations regarding scale and remote sensing. Nevertheless, in general terms, best practice in remote sensing remains, as far as possible, to match project objectives (in terms of the features of interest) with image capabilities (as determined by the spatial resolution(s)) (Marceau et al. 1994, Sawaya et al. 2003, Aplin 2005).

3. Temporal variability

The temporal variability of the Earth's surface is represented in remote sensing according to certain temporal properties of the imagery, in particular the time and date of acquiring images, and the interval between acquiring images at a constant site. This is synonymous with temporal resolution, which refers specifically to the period of time between successive image acquisitions at the same site (Mather 2004). Other, related terms used fairly widely in connection with temporal resolution include the ‘repeat cycle’ or the ‘revisit time’ of remote sensing instruments (i.e. the time taken for the sensor's area of view to return to a specific point on the Earth's surface). Commonly, temporal resolution is reported in reference to space‐borne remote sensing instruments, since these tend to have a fixed orbit and a predictable repeat cycle. On the contrary, airborne remote sensing tends to involve much less regular temporal resolutions, since image acquisition depends on occasional and often irregularly timed flights. However, the temporal resolution of satellite sensors is still far from straightforward, since this depends on various other factors, including the satellite's orbital altitude and the sensor's angle of view (which together influence the area of coverage (swath width) of the image), the instrument's tilting capabilities and the latitude of the area of interest. For example, since wide swath sensors cover relatively large areas, they tend to have a shorter repeat cycle than narrow swath sensors covering small areas. Also, the polar orbit of most satellite sensors means that repeat cycles tend to be shorter at higher latitude locations (e.g. polar regions) than lower latitude locations (e.g. equatorial regions).

Temporal resolution is significant where imagery is used to monitor changing environmental conditions over time. This involves multi‐temporal analysis, whereby, for a given location or area, multiple images are acquired at different times and combined for analysis. Monitoring may occur over a range of temporal scales and the scale used is generally related to the purpose of investigation. For example, weather patterns are relatively dynamic, such that it may be necessary to acquire and analyse images as frequently as every half hour (Mukherjee and Acton 2002, Casanova et al. 2005, Feidas and Cartalis 2005). Alternatively, monitoring crops throughout the growing season may involve acquiring images approximately once a week or once a fortnight (Pinter et al. 2003, Del Frate et al. 2004, Doraiswamy et al. 2004, Granados‐Ramirez et al. 2004, Xiao et al. 2005). Other applications on much longer time‐scales include measuring the rate of deforestation on an annual basis (Shimabukuro et al. 1999, Sgrenzaroli et al. 2002, Greenberg et al. 2005, Ingram and Dawson 2005) or glacial advance over a period of decades (Mennis and Fountain 2001, Paul 2002, Silverio and Jaquet 2005).

Accurate monitoring relies on certain pre‐processing procedures to ensure a valid comparison between multi‐temporal images. First, all remotely sensed images are subject to geometric distortion (Du et al. 2002, Guienko 2004) and, while this may be relatively inconsequential for single‐date analysis where general spatial patterns are sought, it can be a major difficulty for multi‐temporal analysis where images are compared on a spatial basis. Therefore, all investigations involving multi‐temporal imagery require accurate geometric co‐registration to register multiple images to each other or to a common coordinate system (Coppin and Bauer 1996, Eugenio and Marques 2003, Siqueira et al. 2004). Misregistration between images will mean that different areas are being compared over time, leading to errors in analysis (Townshend et al. 1992, Kardoulas et al. 1996, Tucker et al. 2004). Secondly, remotely sensed images can be affected by certain radiometric, atmospheric and illumination distortions. Specifically, differences in instrument calibration, atmospheric conditions and solar and sensor viewing angles can lead to discrepancies between images, whereby the images are not merely representing the reflectance properties of features on the Earth's surface, but include image‐specific radiometric, atmospheric and illumination effects. Clearly, any such inconsistencies between images may create difficulties for multi‐temporal analysis, and corrections are often necessary prior to temporal comparison (Song et al. 2001, Liang et al. 2002, Vermote et al. 2002, Richter and Schlapfer 2002, Steven et al. 2003). Rather than performing absolute corrections to generate true reflectance values, multi‐temporal studies often use image matching or normalization, whereby discrepancies between images are simply averaged out (Du et al. 2002, Canty et al. 2004, Nelson et al. 2005). However, there are certain circumstances where multi‐temporal analysis may not require any radiometric, atmospheric or illumination corrections. For example, such corrections may be irrelevant for multi‐temporal land‐cover classification, since independently classified images rather than original (spectral) images are being compared.

One form of multi‐temporal analysis is change detection. This involves a direct comparison between images to identify the location of change, and to describe that change in terms of its nature (e.g. a changed land‐cover class) and/or magnitude (e.g. an increase or decrease in vegetation index value). Change detection analysis is used in a wide range of applications, including urban development (Zhang et al. 2002, Yang et al. 2003, Cova et al. 2004), coral reef monitoring (Palandro et al. 2003, Philipson and Lindell 2003, Elvidge et al. 2004) and disaster management in response to both natural events, such as landslides (Cheng et al. 2004, Lin et al. 2005, Nichol and Wong 2005), and anthropogenic events, such as armed conflicts (Abuelgasim et al. 1999, Al‐Khudhairy et al. 2005). Various change detection techniques have been employed successfully using remotely sensed imagery (Singh 1989, Coppin et al. 2004). For example, straightforward image differencing, whereby one image is subtracted from another, has been used widely (Dymond et al. 2002, Lu et al. 2004, Jin and Sader 2005, Lin et al. 2005). Other, more complex approaches include post‐classification comparison, where images are classified independently before being compared (Serra et al. 2003, Nichol and Wong 2005); change vector analysis, where change is modelled in spectral and temporal space (Chen et al. 2003, Warner 2005); and composite multi‐temporal analysis, where images are combined to form a single dataset before, for example, land‐cover classification is performed to identify ‘change’ and ‘no change’ classes (Duchemin et al. 1999, Coppin et al. 2004).

As with the tentative nature of recommendations regarding spatial scale and remote sensing, recommendations regarding temporal scale and remote sensing are similarly general. That is, generic or absolute guidelines regarding the temporal resolution of remotely sensed imagery to be used in remote sensing studies are unlikely to be appropriate in all cases. Instead, simply, to increase the likelihood of success of any project, the project's objectives should be matched with image capabilities, in this case determined by the temporal resolution. A simple example of this is to ensure that, where monitoring is required, the temporal resolution of the imagery used is sufficiently frequent to meet the needs of the project. For example, throughout the 1980s and 1990s, the Landsat Thematic Mapper (TM) instrument series was a key source of imagery, used widely in multi‐temporal studies (e.g. Colwell and Poulton 1985, Currey et al. 1987, Pattiaratchi et al. 1994, Guerra et al. 1998). However, Landsat sensors acquired imagery of areas such as the UK only every sixteen days, and the presence of cloud cover meant that perhaps only one in six attempts at acquisition was successful (Legg 1991, Marshall et al. 1994). Such imagery, therefore, was perhaps suitable for monitoring on a quarterly or annual basis, but not on a weekly or daily basis. In contrast, current sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS) have a repeat cycle of only one or two days (Justice et al. 2002), providing data suitable for monitoring on a weekly basis or even more frequently (Doraiswamy et al. 2004, Kang et al. 2005).

4. Reconciling space and time

Clearly, all remote sensing investigations involve a consideration of space. This is implicit in the way remotely sensed data are acquired and presented. Less obvious, perhaps, is that time is also a concern of ‘all’ remote sensing investigations. While some remote sensing analysis is explicitly temporal, whereby phenomena are monitored over time using multi‐temporal imagery, some is not, and the temporal nature of such analysis may be overlooked. This is often the case where static, single‐date imagery is used and initial use of such imagery has no immediate temporal implications. For example, where a land agency has a pressing need for an inventory of resources, they may use remote sensing to create a land‐cover map. While the map may provide a sufficiently accurate representation of land resources at the time and to meet the immediate need, later re‐use of the map may be ill‐advised since the true pattern of land cover may have altered. That is, initially, the temporal nature of the map is essentially inconsequential, given its immediate use. Later, though, any use of the map must consider its historical (i.e. temporal) context, since this will influence its utility.

While space and time are both key elements of remote sensing, to some extent the practice of acquiring remotely sensed imagery involves a play‐off between these two properties. Any remotely sensed image pixel requires a certain amount of computer storage space, regardless of its spatial resolution. Therefore, a fine spatial resolution image covering a relatively small part of the Earth's surface (narrow swath) requires the same storage space as a coarser spatial resolution image covering a larger area (wide swath), providing the two images have the same number of pixels (rows and columns) and other things (e.g. spectral characteristics, radiometric resolution) are equal. Where two such fine and coarse spatial resolution satellite sensors share orbital and other characteristics, the coarse spatial resolution instrument will have the shorter repeat cycle due to its larger area of coverage (Steven 1993). This is a realistic comparison since limits to the computer storage space and processing capabilities of remote sensing instruments mean that it is simply not practical to maintain both a wide swath and a fine spatial resolution.

The play‐off between spatial and temporal resolution can be illustrated with reference to real remote sensing instruments. Consider, for example, the two sensors discussed at the end of the previous section, Landsat TM and MODIS. Landsat TM had a spatial resolution of 30 m (in multi‐spectral mode), a swath width of 185 km and a temporal resolution of 16 days (Mika 1997, Goward et al. 2001, Lee et al. 2004, Markham et al. 2004). In contrast, MODIS has a spatial resolution of between 250 m and 1 km (depending on spectral mode), a swath width of 2330 km and a temporal resolution of 1–2 days (Justice et al. 2002, Barnes et al. 2003). As a more recently developed instrument, MODIS has certain advancements compared to Landsat TM, including enhanced radiometric resolution and computer storage capabilities. Nonetheless, the spatial‐temporal play‐off is clearly evident, with the considerably finer spatial resolution of Landsat TM and the considerably more frequent temporal resolution of MODIS.

Certain other factors influence the relationship between the spatial and temporal properties of remote sensing instruments. Of specific relevance here is a sensor's angle of view and, in particular, its tilting capabilities. The temporal resolution of a satellite sensor can be more frequent where a sensor is able to tilt away from the normal vertically downward (nadir) view of observation to target specific locations of interest on the Earth's surface. That is, while a nadir viewing sensor must wait for the satellite orbit to return to a location, a tilting sensor can point to locations away from the nadir view, essentially increasing the potentially viewable area. However, it is important to realize that selective targeting (e.g. tilting) to increase the temporal resolution at one location automatically means that alternative (e.g. nadir view) targets are deselected.

Many modern satellite sensors have tilting capabilities, but of particular interest here is the generation of new commercial fine spatial resolution satellite sensors (Aplin et al. 1997, Tanaka and Sugimura 2001, Aplin 2003, Birk et al. 2003), including IKONOS, QuickBird and OrbView‐3. These instruments are significant since, although they have very fine spatial resolutions of between 0.6 m and 4 m, they can have temporal resolutions as frequent as 1–2 days due to their flexible tilting capabilities (up to 45° away from nadir) (Dial et al. 2003, Toutin 2004). It should be clear that a sensor's temporal resolution becomes more frequent as its angle of tilt increases, but it is important to note that tilting, particular at high angles, can degrade image quality. Often, therefore, tilting is limited to relatively shallow angles, such that effective temporal resolutions may be less frequent than the optimum (i.e. most frequent) values stated by the supplier or in the literature. Regardless, a comparison between the spatial and temporal characteristics of IKONOS (1 m spatial resolution, 1 day temporal resolution) and Landsat TM (30 m spatial resolution, 16 day temporal resolution) demonstrates the benefits of tilting clearly, although it should be noted that certain other technical factors also contribute to these differences (Goward et al. 2003).

Much reference is made in the literature to ‘spatio‐temporal’ studies in remote sensing (e.g. Jeon and Landgrebe 1992, Herold et al. 2003, Palacios‐Orueta et al. 2004, Benie et al. 2005). Elsewhere, investigation of spatial ‘and’ temporal patterns or relationships are described specifically (e.g. Saitoh et al. 2002, Pavelsky and Smith 2004, Nezlin and Stein 2005, Zarco‐Tejada et al. 2005). In essence, such studies represent a form of multi‐temporal analysis, but spatio‐temporal investigation can extend well beyond simple monitoring. For example, Woodcock et al. (2001) described the transferability of classification techniques across spatial and temporal scales, Lu and Inamura (2003) presented a new technique of super‐resolution mapping that exploits sub‐pixel differences between multi‐temporal images, while Pulliainen et al. (2004) combined multi‐temporal image and surface data with spatial modelling techniques to assess and predict water quality distributions. These three specific topics are examined further in this Special Issue. More generally, spatio‐temporal investigation has been used in a wide range of applications, including, for example, soil moisture analysis (Kim and Barros 2002, Kelly et al. 2003), forest management (Song and Woodcock 2002, Ferreira et al. 2005) and urban growth (Herold et al. 2003, Pellizzeri et al. 2003, Nichol 2005).

5. New developments

The theme ‘Scales and Dynamics in Observing the Environment’ is sufficiently broad to render it impractical to provide a full description of recent developments in a single journal issue. Instead, this Special Issue introduces various specific and noteworthy topics that exemplify emerging trends in this area of investigation.

The cover paper, by Priestnall and Aplin (2006) illustrates spatial and temporal remote sensing requirements for river monitoring. Following this review, the next three papers, by Archer and Jones (2006), Donoghue and Watt (2006) and Mehner et al. (2006), focus principally on issues of space and scale‐dependency in characterizing environmental distributions from remotely sensed imagery. Archer and Jones (2006) investigate problems related to the estimation of surface temperature of heterogeneous vegetation at multiple spatial scales using combined hyperspectral and thermal remotely sensed imagery. Donoghue and Watt (2006) combine aerial photography, IKONOS, Landsat Enhanced Thematic Mapper Plus (ETM+) and Light Detection and Ranging (LiDAR) data to estimate forest height. Rather surprisingly, the finer spatial resolution of IKONOS does not lead to notably more accurate results than Landsat ETM+. Mehner et al. (2006), after establishing an optimum spatial resolution to classify upland vegetation, attempt to transfer or generalize these classification techniques spatially to other, unseen areas.

Transferability also forms the focus of a study by Boyd et al. (2006), but this involves temporal as well as spatial considerations. Specifically, a drought monitoring system is developed, involving a comparison of remotely sensed image products and rainfall measurements, to assess the impact of the El Niño Southern Oscillation (ENSO) on rainforests in Borneo. ENSO is also one of several factors influencing fire occurrence in Bolivia and Peru, investigated by Bradley and Millington (2006). This paper considers spatial and temporal scale issues in the occurrence and observation of fires broadly, comparing various different remote sensing‐derived fire data products.

Murwira and Skidmore (2006) present a new geostatistical approach to monitor change in the spatial heterogeneity of vegetation. This approach, based on measures of intensity and dominant scale of spatial heterogeneity, was used to estimate vegetal change in Zimbabwe from 1984 to 1992. Muslim et al. (2006) also develop a geostatistical technique, in this case to map the position of the shoreline in Malaysia using IKONOS imagery. Specifically, super‐resolution mapping is used to locate the land–sea boundary, a particularly dynamic phenomenon in Malaysia due to severe and rapid erosion and other factors, at the sub‐pixel scale.

The final two papers, by Miller et al. (2006) and Doxaran et al. (2006), focus on scaling issues in observing dynamic marine distributions. Miller et al. (2006) use Sea‐viewing Wide Field‐of‐view Sensor (SeaWiFS) imagery to characterize the spatial evolution of harmful algal blooms. However, while the method developed is relatively effective for large area studies, a multi‐scale approach is recommended, whereby coarse spatial resolution imagery in the open ocean (as provided by SeaWiFS) is supplemented by finer spatial resolution imagery in coastal areas where environmental impacts are most significant. Finally, Doxaran et al. (2006) estimate movements of the maximum turbidity zone in the Gironde estuary in France. Turbidity features are monitored over multiple spatial and temporal scales, including long‐term seasonal and tidal movements throughout the entire estuary, and short‐term shorebank re‐suspension and turbulent currents.